Download Basic_Statistics

BASIC STATISTICS
1. S AMPLES , R ANDOM S AMPLING
AND
S AMPLE S TATISTICS
1.1. Random Sample. The random variables X1 , X2, ..., Xn are called a random sample of size n
from the population f(x) if X1 , X2, ..., Xn are mutually independent random variables and the marginal probability density function of each Xi is the same function of f(x). Alternatively,
X1 , X2, ..., Xn are called independent and identically distributed random variables with pdf f(x).
We abbreviate independent and identically distributed as iid.
Most experiments involve n >1 repeated observations on a particular variable, the first observation is X1 , the second is X2 , and so on. Each Xi is an observation on the same variable and each Xi
has a marginal distribution given by f(x). Given that the observations are collected in such a way
that the value of one observation has no effect or relationship with any of the other observations,
the X1 , X2, ..., Xn are mutually independent. Therefore we can write the joint probability density
for the sample X1 , X2, ..., Xn as
f(x1 , x2, ..., xn) = f(x1 )f(x2 ) · · · f(xn ) =
n
Y
f(xi )
(1)
i=1
If the underlying probability model is parameterized by θ, then we can also write
f(x1 , x2, ..., xn|θ) =
n
Y
(2)
f(xi |θ)
i=1
Note that the same θ is used in each term of the product, or in each marginal density. A different
value of θ would lead to a different properties for the random sample.
1.2. Statistics. Let X1 , X2, ..., Xn be a random sample of size n from a population and let
T (x1 , x2, ..., xn) be a real valued or vector valued function whose domain includes the sample space
of (X1 , X2, ..., Xn). Then the random variable or random vector Y = (X1 , X2 , ..., Xn) is called a
statistic. A statistic is a map from the sample space of (X1 , X2 , ..., Xn) call it X, to some space of
values, usually R1 or Rn. T is what we compute when we observe the random variable X take on
some specific values in a sample. The probability distribution of a statistic Y = T(X) is called the
sampling distribution of Y. Notice that T(·) is a function of sample values only, it does not depend
on any underlying parameters, θ.
1.3. Some Commonly Used Statistics.
1.3.1. Sample mean. The sample mean is the arithmetic average of the values in a random sample.
It is usually denoted
X1 + X2 + ... + Xn
1X
Xi
=
n
n i=1
n
X̄(X1 , X2 , · · ·, Xn) =
The observed value of X̄ in any sample is demoted by the lower case letter, i.e., x̄.
Date: September 27, 2004.
1
(3)
2
BASIC STATISTICS
1.3.2. Sample variance. The sample variance is the statistic defined by
1 X
(Xi − X̄)2
n−1
n
S 2 (X1 , X2 , · · ·, Xn) =
(4)
i=1
The observed value of S 2 in any sample is demoted by the lower case letter, i.e., s2 .
1.3.3. Sample standard deviation. The sample standard deviation is the statistic defined by
√
S = S2
(5)
1.3.4. Sample midrange. The sample mid-range is the statistic defined by
max(X1 , X2 , · · ·, Xn) − min(X1 , X2 , · · ·, Xn)
2
(6)
1.3.5. Empirical distribution function. The empirical distribution function is defined by
1X
I(Xi < x)
n
n
F̂ (X1 , X2, · · ·, Xn)(x) =
(7)
i=1
where F̂ (X1 , X2, ···, Xn)(x) means we are evaluating the statistic F̂ (X1 , X2 , ···, Xn) at the particular
value x. The random sample X1 , X2 , ..., Xn is assumed to come from a probability defined on R1
and I(A) is the indicator of the event A. This statistic takes values in the set of all distribution
functions on R1. It estimates the function valued parameter F defined by its evaluation at x ∈ R1
(8)
F (P )(x) = P [X < x]
2. D ISTRIBUTION
OF
S AMPLE S TATISTICS
2.1. Theorem 1 on squared deviations and sample variances.
Theorem 1. Let x1, x2, · · ·xn be any numbers and let x̄ =
hold.
Pn
Pn
2
2
a: mina i=1 (xP
i − a) =
i − x̄)
i=1 (xP
n
n
2
2
b: (n − 1)s = i=1 (xi − x̄) = i=1 x2i − nx̄2
x1 +x2 +...+xn
.
n
Then the following two items
Part a says that the sample mean is the value about which the sum of squared deviations is
minimized. Part b is a simple identity that will prove immensely useful in dealing with statistical
data.
Proof. First consider part a of theorem 1. Add and subtract x̄ from the expression on the lefthand
side in part a and then expand as follows
n
X
(xi − x̄ + x̄ − a)2 =
i=1
n
X
(xi − x̄)2 + 2
i=1
n
X
(xi − x̄)(x̄ − a) +
i=1
n
X
(x̄ − a)2
(9)
i=1
Now write out the middle term in 9 and simplify
n
X
i=1
(xi − x̄)(x̄ − a) = x̄
n
X
i=1
2
xi − a
n
X
i=1
xi − x̄
n
X
i=1
= nx̄ − anx̄ − nx̄2 + nx̄a
=0
x̄ + x̄
n
X
i=1
a
(10)
BASIC STATISTICS
We can then write 9 as
n
X
2
(xi − a) =
i=1
n
X
2
(xi − x̄) +
3
n
X
i=1
(x̄ − a)2
(11)
i=1
Equation 11 is clearly minimized when a = x̄. Now consider part b of theorem 1. Expand the
second expression in part b and simplify
n
X
(xi − x̄)2 =
n
X
=
i=1
n
X
=
n
X
i=1
x2i − 2x̄
n
X
xi +
i=1
n
X
x̄2
(12)
i=1
x2i − 2nx̄2 + nx̄2
i=1
x2i − nx̄2
i=1
2.2. Theorem 2 on expected values and variances of sums.
Theorem 2. Let X1 , X2 , · · · Xn be a random sample from a population and let g(x) be a function such that
Eg(X1 ) and V arg(X1 )exist. Then following two items hold.
Pn
a: E( i=1
Png(Xi )) = n(Eg(X1 ))
b: V ar( i=1 g(Xi )) = n(V arg(X1 ))
Proof. First consider part a of theorem 2. Write the expected value of the sum as the sum of the
expected values and then note that Eg(X1 ) = Eg(X2 ) = ...Eg(Xi) = ...Eg(Xn) because the Xi are
all from the same distribution.
!
n
n
X
X
E
g(Xi ) =
E(g(Xi )) = n(Eg(X1 ))
(13)
i=1
i=1
First consider part b of theorem 2. Write the definition of the variance for a variable z as E(z−E(z))2
and then combine terms in the summation sign.
!
" n
!#2
n
n
X
X
X
V ar
g(Xi ) = E
g(Xi ) − E
g(Xi )
(14)
i=1
i=1
i=1
Now write out the bottom expression in equation 14 as follows
!
n
X
2
V ar
g(Xi ) = E [g(X1 ) − E(g(X1 ))] + E [g(X1 ) − E(g(X1 ))] E [g(X2 ) − E(g(X2 ))]
i=1
+ E [g(X1 ) − E(g(X1 ))] E [g(X3 ) − E(g(X3 ))] + · · ·
+ E [g(X2 ) − E(g(X2 ))] E [g(X1 ) − E(g(X1 ))] + E [g(X2 ) − E(g(X2 ))]
2
(15)
+ E [g(X2 ) − E(g(X2 ))] E [g(X3 ) − E(g(X3 ))] + · · ·
+· · ·
+ E [g(Xn ) − E(g(Xn ))] E [g(X1 ) − E(g(X1 ))] + · · · + E [g(Xn ) − E(g(Xn ))]2
Each of the squared terms in the summation is a variance, i.e., the variance of Xi = var(X1 ). Specifically
2
E [g(Xi ) − E(g(Xi ))] = V arg(Xi ) = V arg(X1 )
(16)
4
BASIC STATISTICS
The other terms in the summation in 15 are covariances of the form
E [g(Xi ) − E(g(Xi ))] E [g(Xj ) − E(g(Xj ))] = Cov [g(Xi ), g(Xj )]
(17)
Now we can use the fact that the X1 and Xj in the sample X1 , X2, · · ·, Xn are independent to assert
that each of the covariances inn the sum in 15 is zero. We can then rewrite 15 as
!
n
X
2
2
2
g(Xi ) = E [g(X1 ) − E(g(X1 ))] + E [g(X2 ) − E(g(X2 ))] + · · · + E [g(Xn ) − E(g(Xn ))]
V ar
i=1
(18)
= V ar(g(X1 )) + V ar(g(X2 )) + V ar(g(X3 )) + · · ·
n
X
=
V arg(Xi )
i=1
=
n
X
V arg(X1 )
i=1
= n V arg(X1 )
2.3. Theorem 3 on expected values of sample statistics.
Theorem 3. Let X1 , X2, · · ·Xn be a random sample from a population with mean µ and variance σ2 < ∞.
Then
a: E X̄ = µ
2
b: V arX̄ = σn
c: ES 2 = σ2
Proof of part a. In theorem 2 let g(X) = g(Xi ) =
n
E X̄ = E
1X
Xi
n
!
i=1
Proof of part b.
In theorem 2 let g(X) = g(Xi ) =
Xi
n .
n
V arX̄ = V ar
Xi
n .
1
= E
n
This implies that Eg(Xi ) =
n
X
Xi
!
=
i=1
1X
Xi
n i=1
=
n
X
1
V
ar
Xi
n2
i=1
!
Then we can write
1
(nEX1) = µ
n
This implies that V arg(Xi ) =
!
µ
n
=
σ2
n
(19)
Then we can write
1
σ2
(nV
arX
)
=
1
n2
n
(20)
Proof of part c.
As in part b of theorem 1, write S 2 as a function of the sum of square of Xi minus n times the mean of
Xi squared and then simplify
2
ES = E
#!
" n
X
1
2
2
Xi − nX̄
n−1
i=1
1
=
nEX12 − nE X̄ 2
n−1
2
1
σ
=
= σ2
n(σ2 + µ2) − n
+ µ2
n−1
n
The last line follows from the definition of a random variable, i.e.,
(21)
BASIC STATISTICS
5
2
V ar X = σX
= EX 2 − (EX)2
= EX 2 − µ2X
2
⇒EX =
2
σX
−
(22)
µ2X
2.4. Unbiased Statistics. We say that a statistic T(X)is an unbiased statistic for the parameter θ of
the underlying probability distribution if E T(X) = θ. Given this definition, X̄ is an unbiased statistic
for µ,and S 2 is an unbiased statistic for σ2 in a random sample.
3. M ETHODS
OF
E STIMATION
Let Y1 , Y2, · · · Yn denote a random sample from a parent population characterized by the parameters θ1 , θ2, · · · θk . It is assumed that the random variable Y has an associated density function
f( · ; θ1, θ2 , · · · θk ).
3.1. Method of Moments.
3.1.1. Definition of Moments. If Y is a random variable, the rth moment of Y, usually denoted by µ0r ,
is defined as
µ0r = E(Y r )
Z ∞
=
yr f(y; θ1 , θ2 , · · ·θk ) dy
(23)
−∞
if the expectation exists. Note that µ01 = E(Y ) = µY , the mean of Y. Moments are sometimes
written as functions of θ.
E(Y r ) = µ0r = gr (θ1 , θ2 , · · ·θk )
(24)
3.1.2. Definition of Central Moments. If Y is a random variable, the rth central moment of Y about a
is defined as E[(Y − a)r ]. If a = µr , we have the rth central moment of Y about µY , denoted by µr ,
which is
µr = E[(Y − µY )r ]
Z ∞
=
(y − µy )r f(y; θ1 , θ2, · · · θk ) dy
(25)
−∞
Note thatµ1 = E[(Y − µY )] = 0 and µ2 = E[(Y − µY )2 ] = V ar[Y ]. Also note that all odd
numbered moments of Y around its mean are zero for symmetrical dsitributions, provided such
moments exist.
3.1.3. Sample Moments about the Origin. The rth sample moment about the origin is defined as
n
µ̂0r = x̄rn =
1X r
y
n i=1 i
(26)
6
BASIC STATISTICS
3.1.4. Estimation Using the Method of Moments. In general µ0r will be a known function of the parameters θ1 , θ2, · · · θk of the distribution of Y, that is µ0r = gr (θ1 , θ2, · · · θk ). Now let y1 , y2 , · · · , yn be a
random sample from the density f(·; θ1 , θ2, · · · θk ). Form the K equations
1X
=
yi
n
n
µ01
=g1(θ1 , θ2, · · · θk ) =
µ̂01
i=1
n
1X 2
µ02 =g2(θ1 , θ2, · · · θk ) = µ̂02 =
y
n i=1 i
(27)
..
.
n
µ0K =gK (θ1 , θ2, · · · θk ) = µ̂0K =
1X K
yi
n
i=1
The estimators of θ1 , θ2 , · · · θk , based on the method of moments, are obtained by solving the
system of equations for the K parameter estimates θ̂1 , θˆ2 , · · · θˆK .
This principle of estimation is based upon the convention of picking the estimators of θı in such a
manner that the corresponding population (theoretical) moments are equal to the sample moments.
These estimators are consistent under fairly general regularity conditions, but are not generally
efficient. Method of moments estimators may also not be unique.
3.1.5. Example using density function f(y) = (p + 1) yp . Consider a density function given by
f(y) = (p + 1) yp 0 ≤ y ≤ 1
(28)
= 0 otherwise
Let Y1 , Y2, · · · Yn denote a random sample from the given population. Express the first moment
of Y as a function of the parameters.
E(Y ) =
=
Z
∞
y f(y) dy
−∞
Z 1
y (p + 1) yp dy
Z
0
1
yp+1 (p + 1) dy
0
1
yp+2 (p + 1) =
(p + 2) =
0
p+1
=
p+2
Then set this expression of the parameters equal to the first sample moment and solve for p.
(29)
BASIC STATISTICS
µ01 = E(Y ) =
7
p+1
p+2
1X
p+1
yi = ȳ
=
⇒
p+2
n
n
(30)
i=1
⇒ p + 1 = (p + 2) ȳ = pȳ + 2ȳ
⇒ p − pȳ = 2 ȳ − 1
⇒ p(1 − ȳ) = 2 ȳ − 1
2 ȳ − 1
⇒ p̂ =
1 − ȳ
3.1.6. Example using the Normal Distribution. Let Y1 , Y2, · · · Yn denote a random sample from a normal distribution with mean µ and variance σ2 . Let (θ1 , θ2 ) = (µ, σ2). Remember that µ = µ01 and
σ2 = E[Y 2 ] − E 2 [Y ] = µ02 − (µ01)2 .
µ01 = E(Y ) = µ
(31)
µ02 = E(Y 2) = σ2 + E 2 [Y ] = σ2 + µ2
Now set the first population moment equal to its sample analogue to obtain
1X
µ=
yi = ȳ
n
n
(32)
i=1
⇒ µ̂ = ȳ
Now set the second population moment equal to its sample analogue
1X 2
yi
n
n
σ 2 + µ2 =
i=1
n
1X 2
⇒ σ2 =
yi − µ2
n
i=1
v
u n
u1 X
⇒σ=t
yi2 − µ2
n
(33)
i=1
Now replace µ in equation 33 with its estimator from equation 32 to obtain
v
u n
u1 X
σ̂ = t
yi2 − ȳ2
n
i=1
v
u n
uX (yi − ȳ)2
⇒ σ̂ = t
n
i=1
This is, of course, from the sample standard deviation defined in equations 4 and 5.
(34)
8
BASIC STATISTICS
3.1.7. Example using the Gamma Distribution. Let X1 , X2, · · · Xn denote a random sample from a
gamma distribution with parameters θ and α. The density function is given by
−x
1
xα−1 e θ
θα Γ(α)
=0
otherwise
f(x; θ, α) =
0≤x<∞
(35)
Find the first moment of the gamma distribution by integrating as follows
Z
∞
−x
1
xα−1 e θ dx
α Γ(α)
θ
0
Z ∞
−x
1
= α
x(1+α)−1 e θ dx
θ Γ(α) 0
E(X) =
x
(36)
If we multiply equation 36 by θ1+α Γ(1 + α) we obtain
θ1+α Γ(1 + α)
E(X) =
θα Γ(α)
Z
∞
0
−x
1
x(1+α)−1 e θ dx
θ1+α Γ(1 + α)
(37)
The integrand of equation 37 is a gamma density with parameters θ and 1 − α This integrand
will integrate to one so that we obtain the expression in front of the integral sign as the E(X).
θ1+α Γ(1 + α)
θα Γ(α)
θ Γ(1 + α)
=
Γ(α)
(38)
E(X) =
The gamma function has the property that Γ(t) = (t − 1)Γ(t − 1) or Γ(v + 1) = vΓ(v). Replacing
Γ(1 + α) with α Γ(α) in equation 38, we obtain
θ Γ(1 + α)
Γ(α)
θ αΓ(α)
=
Γ(α)
= θα
(39)
E(X) =
We can find the second moment by finding E(X 2 ). To do this we multiply the gamma density
in equation 36 by x2 instead of x. Carrying out the computation we obtain
E(X 2 ) =
Z
∞
x2
0
=
1
θα Γ(α)
1
α
θ Γ(α)
Z
xα−1 e
−x
θ
∞
x(2+α)−1 e
0
dx
−x
θ
dx
(40)
BASIC STATISTICS
9
If we then multiply 40 by θ2+α Γ(2 + α) we obtain
Z
−x
1
θ2+α Γ(2 + α) ∞
x(2+α)−1 e θ dx
E(X 2 ) =
α
2+α
θ Γ(α)
θ
Γ(2 + α)
0
θ2+α Γ(2 + α)
=
θα Γ(α)
θ2 (α + 1) Γ(1 + α)
=
Γ(α)
θ2 α(α + 1) Γ(α)
=
Γ(α)
(41)
=θ2 α(α + 1)
Now set the first population moment equal to the sample analogue to obtain
1X
θα =
xi = x̄
n
n
(42)
i=1
x̄
θ
Now set the second population moment equal to its sample analogue
⇒ α̂ =
n
1 X 2
x
n i=1 i
Pn
2
2
i=1 xi
⇒θ =
n α ( α + 1)
Pn
2
2
i=1 xi
⇒θ =
x̄
+1
n x̄θ
θ
Pn
x2
⇒ θ2 = n x̄ 2 i=1 i n x̄ +
θ2
θ
n
X
⇒n x̄ 2 + n x̄ θ =
x2i
θ2 α ( α + 1 ) =
(43)
i=1
⇒n x̄ θ =
n
X
x2i − n x̄ 2
i=1
Pn
2
2
i=1 xi − n x̄
⇒θ =
n x̄
Pn
2
(x
−
x̄)
i
= i=1
n x̄
3.2. Method of least squares estimation. Consider the situation in which the Yi from the random
sample can be written in the form
Yi = β + i = β̂ + ei
2
(44)
where E(i ) = 0 and Var(i ) = σ for all i. This is equivalent to stating that the population from
which yi is drawn has a mean of β and a variance of σ2.
10
BASIC STATISTICS
The least squares estimator of β is obtained by minimizing the sum of squares errors, SSE, defined by
SSE =
n
X
e2i =
i=1
n
X
(yi − β̂)
2
(45)
i=1
The idea is to pick the value of β̂ to estimate β which minimizes SSE. Pictorially we select the
value of β̂ which minimizes the sum of squares of the vertical deviations in figure 1.
F IGURE 1. Least Squares Estimation
The solution is obtained by finding the value of β that minimizes equation 45.
n
X
∂SSE
(yi − β̂)(−1) = 0
=2
∂β
i=1
⇒ β̂ =
n
1 X
yi = ȳ
n i−1
(46)
This method chooses values of the parameters of the underlying distribution, θ, such that the
distance between the elements of the random sample and “predicted” values are minimized.
BASIC STATISTICS
11
3.3. Method of maximum likelihood estimation (MLE). Least squares is independent of a specification of a density function for the parent population. Now assume that
(47)
yi ∼ f ( · ; θ = ( θ1 , ..., θK ) ) ∀i.
3.3.1. Motivation for the MLE method. If a random variable Y has a probability density function f(·;
θ) characterized by the parameters θ = (θ1 , . . ., θk ), then the maximum likelihood estimators (MLE)
of θ1 , . . ., θk are the values of these parameters which would have most likely generated the given
sample.
3.3.2. Theoretical development of the MLE method. The joint density of a random sample y1, y2, . . .,
yn is given by L = g (y1 , · · · , yn ; θ) = f(y1 ; θ) · f (y2 ; θ) · f(y3 ; θ) · · · · f(yn ; θ) . Given
that we have a random sample, the joint density is just the product of the marginal density functions. This is referred to as the likelihood function. The MLE of the θi are the θi which maximize the
likelihood function.
The necessary conditions for an optimum are:
∂L
= 0 i = 1, 2, ..., k
∂θi
(48)
This gives k equations in k unknowns to solve for the k parameters θ1 , ..., θk . In many instances
it will be convenient to maximize ` = ln L rather than L given that the log of a product is the sum
of the logs.
3.3.3. Example 1. Let the random variable Xi be distributed as a normal N(µ,σ2 ) so that its density
is given by
f(xi ; µ , σ2 ) = √
1
2 π σ2
· e
−1
2
( xi σ− µ )
2
(49)
Its likelihood function is given by
2
L = Πn
i=1 f(xi ; µ , σ ) = f(x1 ) f(x2 ) · · · f(xn )
=
=
√
1
n
e
−1
2
(
xi − µ
σ
)
2
··· e
2πσ2
√
1
n
2πσ2
⇒ ln L = ` =
e
−1
2 σ2
Pn
i=1 (xi
− µ)
−1
2
( xn σ− µ )
2
2
n
−n
1 X
2
(xi − µ )
ln(2 π σ2 ) −
2
2 σ2 i=1
The MLE of µ and σ2 are obtained by taking the partial derivatives of equation 50
(50)
12
BASIC STATISTICS
Pn
n
∂`
1 X
i=1 xi
(xi − µ) = 0 ⇒ µ̂ =
= 2
= x̄
∂µ
σ̂
n
i=1
∂`
−n
=
2
∂σ
2
⇒
2π
2πσ̂2
n
1
−2 X
2
(xi − µ̂) = 0
(−1)(σ̂2 )
2
i=1
−
n
X
n
1
2
=
(xi − µ̂)
2 σ̂2
(2σ̂2 )2
i=1
n
1 X
2
(xi − µ̂)
σ̂2 i=1
⇒n =
⇒ σ̂2 =
(51)
n
1 X
2
(xi − µ̂)
n
i=1
⇒ σ̂2 =
=
n
1 X
2
(xi − x̄)
n i=1
n − 1
n
s2
The MLE of σ2 is equal to the sample variance and not S2; hence, the MLE is not unbiased as can
be seen from equation 21. The MLE of µ is the sample mean.
3.3.4. Example 2 - Poisson. The random variable Xi is distributed as a Poisson if the density of Xi is
given by
f(xi ; λ) =
 −λ
e x
i

λxi
!
xi a non − negative integer
0
otherwise
mean (X) = λ
(52)
V ar (X) = λ
The likelihood function is given by
L =
e−λ λx1
x1 !
· · ·
e−λ λxn
xn !
n
e−λn λΣi=1 xi
=
n x !
πi=1
i
⇒ ln L = ` = −λ n +
(53)
n
X
n
xi ln λ − ln ( πi=1
xi ! )
i=1
To obtain a MLE of λ, differentiate ` with respect to λ:
BASIC STATISTICS
13
X
∂`
1
xi
= −n+
= 0
∂λ
λ
n
i=1
⇒ λ̂ =
Pn
i=1
xi
n
(54)
= x̄
3.3.5. Example 3. Consider the density function
f ( y ) = (p + 1) yp 0 ≤ y ≤ 1
(55)
=0
otherwise
The likelihood function is given by
p
L = Πn
i=1 (p + 1) yi
ln L = ` =
n
X
ln [(p + 1) yip ]
i=1
=
n
X
(56)
(ln (p + 1) + p ln yi )
i=1
To obtain the MLE estimator differentiate 56 with respect to p
n X
∂`
1
=
+ ln yi = 0
∂p
p+1
i=1
⇒
n
X
i=1
⇒
n
X
1
(− ln yi )
=
p̂ + 1
i=1
n
X
n
(− ln yi )
=
p̂ + 1
(57)
i=1
−n
i=1 ln yi
⇒ p̂ + 1 = Pn
−n
−1
i=1 ln yi
⇒ p̂ = Pn
3.3.6. Example 4. Consider the density function
f(y) = pyi (1 − p)1 − yi
The likelihood function is given by
0 ≤ p ≤ 1
(58)
14
BASIC STATISTICS
yi
L = Πn
( 1 − p ) 1 − yi
i=1 p
Pn
yi
ln L = ` =
n
X
= p
i=1
( 1 − p )n −
yi ln p +
Pn
yi
n −
n
X
i=1
i=1
yi
!
(59)
ln (1 − p)
i=1
To obtain the MLE estimator differentiate 59 with respect to p where we assume that 0 < p < 1.
Pn
Pn
∂`
yi
(n −
i = 1 yi )
= i=1
−
= 0
∂p
p
1 − p
Pn
Pn
(n −
i=1 yi
i = 1 yi )
⇒
=
p
1 − p
⇒
n
X
yi − p
i=1
n
X
yi = n p − p
i=1
⇒
n
X
n
X
yi
(60)
i=1
yi = n p
i=1
⇒
Pn
i=1
n
yi
= p̂
3.4. Principle of Best Linear Unbiased Estimation (BLUE).
3.4.1. Principle of Best Linear Unbiased Estimation. Start with some desired properties and deduce
an estimator satisfying them. For example suppose that we want the estimator to be linear in the
observed random variables. This means that if the observations are y1 , ... , yn , an estimator of θ
must satisfy
θ̂ =
n
X
ai yi
(61)
i=1
where the ai are to be determined.
3.4.2. Some required properties of the estimator (arbitrary).
1: E(θ̂) = θ (unbiased)
2: V ar(θ̂) ≤ V AR(θ̃) (minimum variance) where θ̃ is any other linear combination of the yi
that also produces an unbiased estimator.
3.4.3. Example. Let Y1, Y2, . . ., Yn denote a random sample drawn from a population having a
mean µ and variance σ2 . Now derive the best linear unbiased estimator (BLUE) of µ.
Let the proposed estimator be denoted by θ̂. It is linear so we can write it as follows.
θ̂ =
n
X
ai yi
i=1
If the estimator is to be unbiased, there will be restrictions on the ai . Specifically
(62)
BASIC STATISTICS
U nbiasedness
15
n
X
⇒ E(θ̂) = E
ai yi
!
i=1
=
n
X
ai E(yi )
=
i=1
n
X
ai µ
i=1
n
X
=µ
=>
(63)
ai
i=1
n
X
ai = 1
i=1
Now consider the variance of θ̂.
V ar ( θ̂ ) = V ar
"
n
X
ai yi
#
i=1
=
X
a2i V ar(yi ) + Σ Σi6=j ai aj Cov (yi yj )
=
n
X
a2i σ2
(64)
i=1
because the covariance between yi and yj (i 6= j) is equal to zero due to the fact that the y’s are
drawn from a random sample.
The P
problem of obtaining a BLUE of µ becomes that of minimizing
n
straint i = 1 ai = 1 . This is done by setting up a Lagrangian
L(a, λ) =
n
X
a2i − λ(
i=1
n
X
Pn
i=1
a2i subject to the con-
ai − 1)
(65)
i=1
The necessary conditions for an optimum are
∂L
= 2a1 − λ = 0
∂a1
.
.
.
∂L
= 2an − λ = 0
∂an
n
X
∂L
ai + 1 = 0
= −
∂λ
i=1
The first n equations imply that a1 = a2 = a3 = . . . an so that the last equation implies that
(66)
16
BASIC STATISTICS
n
X
ai − 1 = 0
i=1
⇒ nai − 1 = 0
⇒ nai = 1
1
⇒ ai =
n
n
n
X
1 X
⇒ θ̂ =
ai yi =
yi = ȳ
n i=1
i=1
(67)
Note that equal weights are assigned to each observation.
4. F INITE S AMPLE P ROPERTIES
OF
E STIMATORS
4.1. Introduction to sample properties of estimators. In section 3 we discussed alternative methods of estimating the unknown parameters in a model. In order to compare the estimating techniques we will discuss some criteria which are frequently used in such a comparison. Let θ denote
an unknown parameter and let θ̂ and θ̃ be alternative estimators. Now define the bias, variance and
mean squared error of θ̂ as
Bias (θ̂) = E (θ̂) − θ
V ar (θ̂) = E
M SE (θ̂ ) = E
θ̂ − E (θ̂)
θ̂ − θ
2
= V ar (θ̂) +
2
Bias
(68)
2
θ̂
The result on mean squared error can be seen as follows
2
M SE(θ) = E θ̂ − θ
2
= E θ̂ − E θ̂ + E θ̂ − θ
=E
2
θ̂ − E θ̂
+ E θ̂ − θ
2
h i
2
= E θ̂ − E θ̂
+ 2 E θ̂ − θ E θ̂ − E θ̂
+ E θ̂ − θ
2
2
= E θ̂ − E (θ̂)
+ E (θ̂) − θ
since E θ̂ − E θ̂
= 0
2
= V ar θ̂ + Bias(θ̂)
4.2. Specific properties of estimators.
(69)
BASIC STATISTICS
17
4.2.1. Unbiasedness. θ̂ is said to be an unbiased estimator of θ if E θ̂ = θ .
In figure 2, θ̂ is an unbiased estimator of θ, while θ̃ is a biased estimator.
F IGURE 2. Unbiased Estimator
fHΘL
Ž
fHΘL
ï
fHΘL
Θ
Θ0
4.2.2. Minimum variance. θ̂ is said to be a minimum variance estimator of θ if
V ar θ̂ ≤ V ar θ̃
(70)
where θ̃ is any other estimator of θ. This criterion has its disadvantages as can be seen by noting
that θ̂ = constant has zero variance and yet completely ignores any sample information that we may
have. In figure 3, θ̃ has a lower variance than θ̂.
4.2.3. Mean squared error efficient. θ̂ is said to be a MSE efficient estimator of θ if
M SE θ̂ ≤ M SE θ̃
(71)
where θ̃ is any other estimator of θ. This criterion takes into account both the variance and bias
of the estimator under consideration. Figure 4 shows three alternative estimators of θ.
4.2.4. Best linear unbiased estimators. θ̂ is the best linear unbiased estimator (BLUE) of θ if
18
BASIC STATISTICS
F IGURE 3. Estimators with the Same Mean but Different Variances
fHΘL
Ž
fHΘL
ï
fHΘL
Θ
θ̂ =
n
X
ai yi linear
i=1
E(θ̂) = θ unbiased
(72)
V ar(θ̂) ≤ V ar(θ̃)
where θ̃ is any other linear unbiased estimator of θ.
For the class of unbiased estimators of θ̂, the efficient estimators will also be minimum variance
estimators.
4.2.5. Example. Let X1 , X2 , . . ., Xn denote a random sample drawn from a population having a
population mean equal to µ and a population variance equal to σ2. The sample mean (estimator of
µ) is calculated by the formula
X̄ =
n
X
Xi
n
i=1
(73)
and is an unbiased estimator of µ from theorem 3 and equation 19.
Two possible estimates of the population variance are
σ̂2 =
n
2
X
(Xi − X̄)
n
i=1
S2 =
n
2
X
(Xi − X̄)
n−1
i=1
We have shown previously in theorem 3 and equation 21 that σ̂2 is a biased estimator of σ2;
whereas S2 is an unbiased estimator of σ2 . Note also that
BASIC STATISTICS
19
F IGURE 4. Three Alternative Estimators
fHΘL
Ž
fHΘL
ï
fHΘL
\
fHΘL
Θ
σ̂
2
2
E σ̂
=
=
n − 1
n
n − 1
n
n − 1
n
Also from theorem 3 and equation 20, we have that
=
V ar X̄
=
S2
E S2
σ2
σ2
n
(74)
(75)
Now consider the mean square error of the two estimators X̄ and S2 where X1, X2, . . . Xn are a
random sample from a normal population with a mean of µ and a variance of σ2 .
20
BASIC STATISTICS
E X̄ − µ
2
= V ar X̄
=
σ2
n
(76)
4
2
σ
=
E S2 − σ
= V ar S
n − 1
2
The variance of S was derived in the lecture on sample moments. The variance of σ̂2 is easily
computed given the variance of S2. Specifically,
n − 1 2
V arσ̂2 = V ar
s
n
2
n − 1
=
V ar S 2
n
(77)
2
n − 1
2 σ4
=
n
n − 1
2 (n − 1) σ4
=
n2
2
We can compute the MSE of σ̂ using equations 68, 74, and 77 as follows
2 2
2
2
2 (n − 1) σ4
n − 1
2
2
+
−
σ
σ
n2
n
2
2 (n − 1) σ4
n − 1
n − 1
4
=
+
σ − 2
σ4 + σ4
n2
n
n
2 (n − 1)
(n − 1)2
2 n (n − 1)
n2
= σ4
+
−
+
n2
n2
n2
n2
2 n − 2 + n2 − 2 n + 1 − 2 n2 + 2 n + n2
= σ4
n2
2n − 1
4
=σ
n2
M SE σ̂2 = E σ̂2 − σ
2
=
Now compare the MSE’s of S2 and σ̂2 .
2n − 1
2
2
4
4
M SE σ̂ = σ
< σ
= M SE S 2
n2
n − 1
So σ̂2 is a biased estimator of S2 but has lower mean square error.
(78)
(79)